Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
MixSignGraph: A Sign Sequence is Worth Mixed Graphs of Nodes
Authors: Shiwei Gan, Yafeng Yin, Zhiwei Jiang, Lei Xie, Sanglu Lu, Hongkai Wen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments conducted on the current five sign language datasets demonstrate that Mix Sign Graph surpasses the most current models on multiple sign language tasks across several datasets, without relying on any additional cues. |
| Researcher Affiliation | Academia | State Key Laboratory for Novel Software Technology, Nanjing University, China Department of Computer Science, The University of Warwick, UK EMAIL EMAIL |
| Pseudocode | No | The paper describes the architecture and modules using descriptive text and figures (e.g., Figure 2 and 3) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | Yes | Code and models are available at: https://github.com/gswycf/Sign Language. |
| Open Datasets | Yes | Our experiment evaluation is conducted on five publicly available SL datasets, including PHOENIX14 (25), PHOENIX14T (5),CSL-Daily (14), How2sign (26) and Open ASL (27). |
| Dataset Splits | Yes | PHOENIX14 (25): A widely used German SL dataset with 1295 glosses from 9 signers for CSLR. It includes 5672, 540 and 629 weather forecast samples for training, validation, and testing, respectively. PHOENIX14T (5): ... It contains 7096, 519 and 642 samples from 9 signers for training, validation and testing, respectively. CSL-Daily (14): A Chinese SL dataset with 18401, 1176, 1077 labeled videos from 10 signers for training, testing and validation. ... We provide detailed information of the five datasets in Table 11. |
| Hardware Specification | Yes | In addition, to ensure the model can be trained end-to-end on three 3090 GPUs with 24GB memory, the entire model is trained in half-precision. |
| Software Dependencies | Yes | Our architecture is implemented using Py Torch 1.11. |
| Experiment Setup | Yes | Our architecture is implemented using Py Torch 1.11. The setup includes the following components: (1) Patchify stem and Patch merging: We use a patchify stem and patch merging module in Sign Graph to obtain patch (node) embeddings. (2) Distance function: In the baseline setting, we measure the distance between two nodes using the Euclidean distance. (3) Global feature module: This module comprises two 1D convolution blocks, a 2layer Bi LSTM with a hidden size of 1024 for global feature modeling, and a fully connected layer for the final prediction. (4) Translation network: For a fair comparison, following the current SOTA SLT model (7), we adopt the pretrained m BART model provided by Huggingface as our translation network. Training Setting. For fair comparisons, we adopt the same data preprocessing and training settings as in previous work (2; 4). In regard to How2Sign and Open ASL datasets, we adopt the visual features generated by a pre-trained I3D model (8) . In addition, to ensure the model can be trained end-to-end on three 3090 GPUs with 24GB memory, the entire model is trained in half-precision. ... For example, in Section A.6, the optimal K values (3, 4, 49, 49) are determined for the LSG and TSG modules. |